Instead it generates estimates of the threshold values between rating notches allowing an
assessment of the shape of the ratings curve. Given the data requirements, the method
was only applied to the full sample, which appears appropriate in view of the overall
robustness of the empirical results to the use of sub-samples.
The results from the ordered probit estimation validate the findings highlighted above
(see Tables 13 and 14, respectively for the ordered probit robust standard errors and the
random effects ordered probit). The core variables identified in the linear regressions
also show up with the correct sign in the ordered probit approach. In addition, the
ordered probit models suggest the significance of somewhat more explanatory variables,
namely inflation and the current account, which were significant only in some
specifications in the linear approach. At the same time, in the area of external variables,
reserves do not show up significantly for Moody’s and Fitch in the restricted
specifications, both for the ordered probit and for the random effects ordered probit.
Finally, for the current account variable, the restricted specification for Moody’s shows
a negative sign for deviations from the long-term average, but a positive sign for the
average, and similar sign switches appear also in some instances for the other agencies.
This result goes some way in reconciling the counter intuitive result of the negative
effect of current account on sovereign ratings, with the conventional wisdom.
[Insert Tables 13 and 14 here]
The estimated threshold coefficients reported in the second part of the tables suggest
that the linear specification assumed for the panel regression above is appropriate. The
plot of the results of the random effects ordered probit (see Figure 2), shows that for all
three agencies the thresholds between rating notches are broadly equally distributed
across the ratings range. In other words, the distance for a country to move e.g. from B-
to B is roughly equal to that for moving from AA to AA+. Nevertheless, the
econometric tests at the bottom of the tables reveal additional insights. For the restricted
model of Moody’s, the test does not reject the null hypothesis of equal distances
between thresholds, but the significance level is close to 10 per cent. Indeed the
estimated thresholds point to a relatively large jump between the ratings for BBB- and
BBB. This suggests that countries close to the non-investment grade rating are given a
wider range before they actually cross that threshold. For Fitch, the hypothesis of equal
22